Self-potential (SP) method represents a passive geophysical technique widely applied in mineral exploration, hydrogeophysics, and archaeological prospecting. Traditional interpretation approaches rely on derivative-based optimization methods that frequently encounter difficulties with parameter uniqueness, convergence to local minima, and sensitivity to noise contamination. This study presents a new approach which incorporates physics-informed neural network (PINN) framework for automated and simultaneous multi-anomaly quantitative interpretation. The PINN training objective employs a weighted multi-term loss function consisting of data misfit between neural network predictions and observed measurements, physics-informed residuals enforcing consistency with analytical forward models, and regularization constraints which prevents source position drift. PINN architecture embeds analytical forward models for various geometric sources including inclined sheets, spheres, and horizontal and vertical cylinders directly into the neural network training. Validation against published field data from the Banias archaeological site in Northern Israel demonstrates that the method recovers source parameters (depth, horizontal position of the center of anomaly) with deviations typically less than 5% from independent interpretations, while achieving root-mean-square errors of 0.018–0.038 mV. The framework successfully processes multiple anomalies simultaneously without requiring isolation procedures, addressing a critical limitation of existing interpretation methods. The complete methodology has been implemented as a publicly accessible web-based application, enabling researchers and practitioners worldwide to perform quantitative SP interpretation through an intuitive interface without requiring specialized programming expertise.